Helix 2.5 released; Figure claims zero-shot generalisation across 30 homes

Figure/Helix

Figure says its Helix 2.5 model did three whole-body household jobs (tidying living rooms, folding towels and making beds) in 30 Bay Area homes it had never seen, handling objects absent from its training data. Pretraining on Index, Figure's large dataset of video of people doing everyday tasks, raised zero-shot success from 9% to 56%. Figure's point: learning from human video, not just robot data, is what lets the robot cope with new homes.

• 'Zero-shot' here means the homes and objects were new; the three tasks themselves were ones Helix had been taught.
• 56% success means the robot still fails at roughly half of attempts in a new home.
• Helix 2.5 matched its predecessor's success using half as much robot-specific adaptation data.
• Figure says it predicted its largest training run's test loss before training began, across an 8× range of data: evidence that more data keeps paying off.